A soft measurement method and device for industrial multi-rate acquisition and a medium

Abstract

The application provides a soft measurement method, device and medium for industrial multi-rate acquisition, which comprises the following steps: collecting historical data in an industrial process, and dividing the historical data to obtain a plurality of sampling rate data; respectively extracting features of different sampling rate data, and dimensionally normalizing each scale feature; obtaining each bidirectional cross attention matrix according to each scale feature and a query source, and fusing the bidirectional cross attention matrices to obtain a bidirectional cross attention feature; dynamically calibrating the bidirectional cross attention feature based on a historical feature to obtain a soft measurement model; and inputting to-be-predicted data in the industrial process into the soft measurement model to obtain a quality variable prediction result in the industrial process. The method, device and medium can solve the problem that the existing soft measurement method of an industrial system has insufficient feature extraction, insufficient cross-scale association mining and limited utilization of historical target trend information when processing multi-sampling rate data, thereby resulting in low prediction accuracy of the quality variable in the industrial process.

Description

Technical Field

[0001] This invention relates to the field of quality prediction technology, and in particular to a soft measurement method, device and medium for industrial multi-rate acquisition. Background Technology

[0002] In modern industrial processes, real-time and accurate measurement and monitoring of critical product quality are crucial for effectively tracking production efficiency and maintaining safe production standards. However, critical product quality indicators are often difficult to measure directly in real time due to challenges such as high temperatures and pressures in industrial environments, strong electromagnetic interference, and the high cost and complex maintenance of specialized measuring instruments. Therefore, soft sensor modeling technology has become a core method for indirectly estimating critical quality indicators through easily measurable process variables.

[0003] To fully leverage the potential of industrial big data and meet the ever-changing demands of industrial production, deep learning-based soft sensing methods have been widely applied in industrial quality prediction in recent years. However, the sampling frequencies of various sensors in industrial systems differ significantly, resulting in process data exhibiting typical multi-sampling-rate characteristics. This data characteristic presents significant challenges to industrial quality prediction and soft sensing modeling. Existing quality prediction methods mainly rely on single-sampling-rate data or perform data alignment through interpolation and padding, leading to insufficient feature extraction from multi-sampling-rate data and inadequate cross-scale correlation mining. Furthermore, most methods do not fully utilize historical target trend information during the prediction process, resulting in insufficient modeling ability for time dependencies and long-term trends in industrial processes, thus affecting prediction accuracy and stability.

[0004] In summary, existing soft measurement methods for industrial systems suffer from insufficient feature extraction, inadequate cross-scale correlation mining, and limited utilization of historical target trend information when processing multi-sampling rate data, resulting in low accuracy in predicting industrial process quality variables. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the above-mentioned shortcomings of the prior art by providing a soft measurement method, device and medium for industrial multi-rate data acquisition, so as to solve the problem that the existing soft measurement methods for industrial systems have insufficient feature extraction, insufficient cross-scale correlation mining and limited utilization of historical target trend information when processing multi-sampling rate data, resulting in low accuracy in predicting industrial process quality variables.

[0006] In a first aspect, the present invention provides a soft measurement method for industrial multi-rate data acquisition, comprising:

[0007] Historical data from industrial processes are collected, and the data in the training set corresponding to the historical data are divided according to the sampling frequency to obtain multiple sampling rate data. The historical data includes process variables and quality variables.

[0008] The sampling rate data at different sampling frequencies are extracted independently by multiple independent branches in the parallel multi-scale convolutional neural network module, and the output after independent feature extraction is dimensionally normalized to obtain the scale features of each sampling rate data.

[0009] Using the scale features with the highest and lowest sampling frequencies as query sources, and based on each scale feature and the query source, a bidirectional cross-attention matrix for each scale feature is obtained through a bidirectional cross-attention module. All bidirectional cross-attention matrices are then fused according to the average fusion principle to obtain bidirectional cross-attention features.

[0010] The historical target value sequence is converted into historical features by guiding the attention module through historical target values, and the bidirectional cross-attention features are dynamically calibrated based on the historical features to obtain a soft measurement model, wherein the historical features have the same dimension as the bidirectional cross-attention features;

[0011] The data to be predicted in the industrial process is input into the soft measurement model to obtain the prediction results of the quality variables in the industrial process.

[0012] Furthermore, each of the independent branches sequentially comprises: a first 1D convolutional layer, a first batch normalization layer, a first ReLU activation function, a max pooling layer, a second 1D convolutional layer, a second batch normalization layer, a second ReLU activation function, and an adaptive average pooling layer.

[0013] Further, the step of using the scale features with the highest and lowest sampling frequencies as query sources, and obtaining a bidirectional cross-attention matrix for each scale feature through a bidirectional cross-attention module based on each scale feature and the query source, specifically includes:

[0014] The scale features of all the sampling rate data are sorted from high to low according to the sampling frequency;

[0015] Calculate the highest frequency query value based on the scale feature of the highest sampling frequency;

[0016] Calculate the lowest frequency query value based on the scale characteristics of the lowest sampling frequency;

[0017] Calculate the key and value for each scale feature based on each scale feature;

[0018] Based on the highest and lowest frequency query values, as well as the key and value of each scale feature, a multi-head attention mechanism is used to calculate the bidirectional cross-attention matrix for each scale feature.

[0019] Furthermore, the bidirectional cross-attention matrix includes a high-frequency attention matrix and a low-frequency attention matrix. The step of fusing all the bidirectional cross-attention matrices according to an average fusion principle to obtain bidirectional cross-attention features specifically includes:

[0020] The high-frequency attention matrix and low-frequency attention matrix of each scale feature are fused to obtain the fusion matrix of each scale feature;

[0021] Each of the fusion matrices is concatenated, and the concatenated matrices are then dimensionally compressed to obtain the bidirectional cross-attention features.

[0022] Furthermore, the fusion of the high-frequency attention matrix and the low-frequency attention matrix for each scale feature to obtain a fusion matrix for each scale feature specifically includes:

[0023] The high-frequency attention matrix and low-frequency attention matrix of each scale feature are fused using the following formula:

[0024] ;

[0025] in, Let r be the fusion matrix of the r-th scale feature. Let be the high-frequency attention matrix for the r-th scale feature. Let be the low-frequency attention matrix for the r-th scale feature.

[0026] Furthermore, before the attention module is guided by historical target values ​​to convert the sequence of historical target values ​​into historical features, the method further includes:

[0027] Based on the samples from the previous preset number of time steps corresponding to the current moment, construct the historical target value sequence for the current moment.

[0028] Furthermore, the step of converting the historical target value sequence into historical features through the historical target value-guided attention module, and dynamically calibrating the bidirectional cross-attention features based on the historical features to obtain a soft measurement model, specifically includes:

[0029] The dimension of the bidirectional cross-attention feature is compressed by linear embedding, and the current time-time sequence of historical target values ​​is transformed into historical features with the same dimension as the compressed bidirectional cross-attention feature by linear embedding.

[0030] Calculate the query value corresponding to the historical features based on the historical features;

[0031] Based on the compressed bidirectional cross-attention features, calculate the keys and values ​​of the compressed bidirectional cross-attention features;

[0032] Calculate the attention weight guided by the historical features based on the query value, key, and value.

[0033] Based on the attention weights guided by the historical features, the compressed bidirectional cross-attention features are dynamically calibrated to obtain the feature vector guided by the historical target value sequence, thereby obtaining the soft measurement model.

[0034] Secondly, the present invention provides a soft measurement device for industrial multi-rate data acquisition, comprising:

[0035] The data acquisition and segmentation module is used to collect historical data from industrial processes and segment the data in the training set corresponding to the historical data according to the sampling frequency to obtain multiple sampling rate data. The historical data includes process variables and quality variables.

[0036] An extraction and normalization module, connected to the acquisition and partitioning module, is used to extract independent features from the sampling rate data at different sampling frequencies through multiple independent branches in the parallel multi-scale convolutional neural network module, and to normalize the dimensions of the output after independent feature extraction to obtain the scale features of each sampling rate data.

[0037] A fusion module is obtained and connected to the extraction and normalization module. It is used to take the scale features with the highest and lowest sampling frequencies as query sources, and obtain the bidirectional cross-attention matrix of each scale feature through the bidirectional cross-attention module according to each scale feature and the query source. Then, it fuses all the bidirectional cross-attention matrices according to the average fusion principle to obtain the bidirectional cross-attention feature.

[0038] A conversion calibration module, connected to the fusion module, is used to guide the attention module to convert the historical target value sequence into historical features through historical target values, and to dynamically calibrate the bidirectional cross-attention features based on the historical features to obtain a soft measurement model, wherein the historical features have the same dimension as the bidirectional cross-attention features;

[0039] The input module, connected to the conversion calibration module, is used to input the data to be predicted in the industrial process into the soft measurement model to obtain the prediction results of the quality variables in the industrial process.

[0040] Thirdly, the present invention provides a soft measurement device for industrial multi-rate acquisition, comprising a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to implement the soft measurement method for industrial multi-rate acquisition described in the first aspect above.

[0041] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the soft measurement method for industrial multi-rate acquisition described in the first aspect.

[0042] The present invention provides a soft measurement method, device, and medium for multi-rate data acquisition in industrial applications. First, historical data from the industrial process is collected, and the data in the corresponding training set is divided according to the sampling frequency to obtain multiple sampling rate data. The historical data includes process variables and quality variables. Then, multiple independent branches in a parallel multi-scale convolutional neural network module extract independent features from the sampling rate data at different sampling frequencies, and the outputs after independent feature extraction are dimensionally normalized to obtain scale features for each sampling rate data. Next, using the scale features at the highest and lowest sampling frequencies as query sources, a bidirectional cross-attention module is used to obtain a bidirectional cross-attention matrix for each scale feature based on each scale feature and the query source. All bidirectional cross-attention matrices are then fused according to an average fusion principle to obtain bidirectional cross-attention features. Then, a historical target value-guided attention module converts the historical target value sequence into historical features, and the bidirectional cross-attention features are dynamically calibrated based on these historical features to obtain a soft measurement model. The historical features and the bidirectional cross-attention features have the same dimension. Finally, the data to be predicted in the industrial process is input into the soft measurement model to obtain the prediction results for the quality variables in the industrial process. This invention collects historical data containing process and quality variables, divides the data into multiple sampling rate data based on sampling frequency, laying the data foundation for subsequent multi-scale feature extraction. Through a parallel multi-scale convolutional neural network module, it performs independent feature extraction on data with different sampling frequencies using multiple independent branches. Then, it performs dimensionality regularization on the outputs of each branch to ensure a consistent feature structure, avoiding feature distortion that may occur when a single network processes heterogeneous sampling data. Through a bidirectional cross-attention module, using the highest and lowest frequency feature branches as dual query sources, it mines the bidirectional correlation of full-scale features, obtaining a bidirectional cross-attention matrix for full-scale features, thus ensuring complete... This method covers the dynamic influence and constraints of cross-scale variables and uses the average fusion principle to fuse multiple attention matrices. It can retain the correlation information brought by the most frequent and least frequent queries in a balanced way. By guiding the attention module with historical target values, the historical target value sequence is embedded into a feature vector of the same dimension, and the cross-attention features are dynamically calibrated. This fully utilizes the historical target trend information and ultimately greatly improves the accuracy of industrial system quality prediction. It solves the problem that existing soft measurement methods for industrial systems have insufficient feature extraction, insufficient cross-scale correlation mining, and limited utilization of historical target trend information when processing multi-sampling rate data, resulting in low accuracy of industrial process quality variable prediction. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this drawing or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this drawing. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0044] Figure 1 This is a flowchart of a soft measurement method for industrial multi-rate data acquisition according to Embodiment 1 of the present invention;

[0045] Figure 2 This is a flowchart of another soft measurement method for industrial multi-rate data acquisition according to Embodiment 1 of the present invention;

[0046] Figure 3 This is a diagram of the soft measurement model for multi-scale convolutional bidirectional cross-attention in an embodiment of the present invention.

[0047] Figure 4 This is a process diagram of hydrocracking according to an embodiment of the present invention;

[0048] Figure 5 This is a comparison chart of predicted and actual values ​​in an embodiment of the present invention;

[0049] Figure 6 This is a schematic diagram of the soft measurement device for industrial multi-rate data acquisition according to Embodiment 2 of the present invention;

[0050] Figure 7 This is a schematic diagram of the soft measurement device for industrial multi-rate acquisition according to Embodiment 3 of the present invention. Detailed Implementation

[0051] To enable those skilled in the art to better understand the technical solution of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0052] It is understood that the specific embodiments and accompanying drawings described herein are merely for explaining the invention and are not intended to limit the invention.

[0053] It is understood that, without conflict, the various embodiments and features in the embodiments of the present invention can be combined with each other.

[0054] It is understood that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, while the parts unrelated to the present invention are not shown in the drawings.

[0055] It is understood that each unit or module involved in the embodiments of the present invention may correspond to only one entity structure, or may be composed of multiple entity structures, or multiple units or modules may be integrated into one entity structure.

[0056] It is understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of this invention may occur in a different order than that marked in the accompanying drawings.

[0057] It is understood that the flowcharts and block diagrams of this invention illustrate the possible architecture, functions, and operations of systems, apparatuses, devices, and methods according to various embodiments of this invention. Each block in the flowchart or block diagram may represent a unit, module, program segment, or code, containing executable instructions for implementing the specified function. Furthermore, each block or combination of blocks in the block diagram and flowchart can be implemented using a hardware-based system to achieve the specified function, or using a combination of hardware and computer instructions.

[0058] It is understood that the units and modules involved in the embodiments of the present invention can be implemented by software or by hardware. For example, the units and modules can be located in a processor.

[0059] Example 1:

[0060] This embodiment provides a soft measurement method for industrial multi-rate data acquisition, such as... Figure 1 As shown, the method includes:

[0061] Step S101: Collect historical data from the industrial process, and divide the data in the training set corresponding to the historical data according to the sampling frequency to obtain multiple sampling rate data, wherein the historical data includes process variables and quality variables.

[0062] It should be noted that historical data in industrial processes can refer to time-series data from multiple sensors in a process industry system.

[0063] Specifically, historical data from industrial processes are collected and normalized, and then divided into training and testing sets in chronological order. The historical data includes process variables and quality variables.

[0064] In the training set, process variables with the same sampling frequency are grouped together, and multiple groups are sorted from high to low sampling frequency to obtain multi-sampling rate data G1,…,G r …G R Where G1 represents the combination of the first sampling rate, G r G represents the combination of the r-th sampling rate. R This represents the combination of the Rth sampling rate.

[0065] Step S102: Extract independent features from the sampling rate data at different sampling frequencies through multiple independent branches in the parallel multi-scale convolutional neural network module, and perform dimension normalization on the output after independent feature extraction to obtain the scale features of each sampling rate data.

[0066] Specifically, a parallel multi-scale convolutional neural network module is constructed. The module includes multiple independent branches, each of which performs independent feature extraction for data with different sampling frequencies to extract spatiotemporal dynamic features at each time scale and avoid the problem of time misalignment. The output feature dimensions are kept consistent through dimension mapping (i.e., dimension regularization) to obtain the scale features of each sampling rate data, providing a unified feature representation basis for subsequent multi-scale information collaboration.

[0067] In one optional embodiment, each of the independent branches sequentially comprises: a first 1D convolutional layer, a first batch normalization layer, a first ReLU activation function, a max pooling layer, a second 1D convolutional layer, a second batch normalization layer, a second ReLU activation function, and an adaptive average pooling layer.

[0068] Specifically, the dimension normalization and independent feature extraction processing of each branch is "one-layer 1D convolution + batch normalization + ReLU activation + max pooling + one-layer 1D convolution + batch normalization + ReLU activation + adaptive average pooling", wherein the first layer 1D convolution is used to extract basic temporal features;

[0069] Conv1D f =X f W f + b f ;

[0070] Among them, Conv1D f This is represented as the convolution operation and output. This represents a one-dimensional convolution operation, with a kernel size preferably of 3 and padding preferably of 1 to preserve the time dimension. X f It is the original input feature map, W f It is a convolution kernel, b f It is a bias term.

[0071] Batch Normalization (BN) is used to normalize the output of convolutions, thus avoiding the vanishing gradient problem and accelerating model convergence. The formula for BN is:

[0072] ;

[0073] Where BN represents the batch normalization operation and output, σ is the batch statistic, and γ and β are learnable parameters. 1The output of the previous convolutional layer is also the input for batch normalization.

[0074] Then, the ReLU activation function is used to introduce nonlinearity to fit the nonlinear relationship between the process variable and the quality variable. The formula for the ReLU activation function is:

[0075] ReLU(H 2 )=max(0,H 2 );

[0076] Among them, H 2 This is the output of the previous batch of normalized data.

[0077] Then use max pooling:

[0078] MaxPool1d(H 3 ) i,j = max (H 3 ) i,k K∈[2j,2j+2];

[0079] Wherein, H 3 Let i be the feature map after the previous processing, j be the index of the first channel, k be the index of the output feature map, and K ∈ [2j, 2j+2] be the range of the pooling window.

[0080] Then, a second 1D convolutional layer is used to deepen feature extraction, followed by a batch normalization layer and a ReLU activation function, which follow the same post-processing logic as the first layer. Finally, an adaptive average pooling layer is used to compress the feature map to a single time step.

[0081] ;

[0082] Among them, H 4 Here, b is the output feature map of the previous layer, c is the second channel index, s is the sequence position index, and L is the length of the input sequence. Then, a flattening operation is performed to obtain a single-scale feature vector.

[0083] ;

[0084] Among them, H 5 The output feature map of the adaptive average pooling layer is represented by B, which is the batch size, representing the number of samples in a batch, and C is the number of channels, representing the number of channels in the feature map.

[0085] The parallel multi-scale convolutional neural network module described above is used to extract features from all groups respectively; finally, R groups of multi-scale features are output.

[0086] Step S103: Using the scale features with the highest and lowest sampling frequencies as query sources, based on each scale feature and the query source, obtain the bidirectional cross-attention matrix of each scale feature through the bidirectional cross-attention module, and fuse all the bidirectional cross-attention matrices according to the average fusion principle to obtain the bidirectional cross-attention feature.

[0087] Specifically, a bidirectional cross-attention module is constructed. This module takes the interaction relationship between high / low feature groups as the core of modeling and realizes the bidirectional correlation modeling of multi-scale process features with quality variables through a dual query source mechanism.

[0088] Specifically, the highest sampling frequency group scale feature and the lowest sampling frequency group scale feature are used as dual query sources. Based on each scale feature and the dual query sources, a bidirectional cross-attention matrix for each scale feature is obtained. All bidirectional cross-attention matrices are fused according to the average fusion principle to obtain the bidirectional cross-attention feature.

[0089] In an optional embodiment, the step of using the scale features with the highest and lowest sampling frequencies as query sources, and obtaining a bidirectional cross-attention matrix for each scale feature through a bidirectional cross-attention module based on each scale feature and the query source, specifically includes:

[0090] The scale features of all the sampling rate data are sorted from high to low according to the sampling frequency;

[0091] Calculate the highest frequency query value based on the scale feature of the highest sampling frequency;

[0092] Calculate the lowest frequency query value based on the scale characteristics of the lowest sampling frequency;

[0093] Calculate the key and value for each scale feature based on each scale feature;

[0094] Based on the highest and lowest frequency query values, as well as the key and value of each scale feature, a multi-head attention mechanism is used to calculate the bidirectional cross-attention matrix for each scale feature.

[0095] Specifically, the R-group multi-scale features are sorted from high to low output frequency, and are denoted as: h 1 ,h r ,…,h R Using the highest frequency group scale feature h 1 Calculate the most frequent query value Q 1 Lowest frequency group feature h R Calculate the lowest frequency query value Q R The key (K) and value (V) are derived from all scale features {h 1 ,h r ,…,hR The calculation yielded the result.

[0096] A multi-head attention mechanism is employed, with the highest frequency query value Q being used as the primary focus. 1 Lowest frequency query value Q R For the query value, check all scale features {h} 1 ,h r ,…,h R The key (K) and value (V) of} are used to calculate the attention weights and bidirectional cross-attention matrix for each scale feature.

[0097] The bidirectional cross-attention matrix includes a high-frequency attention matrix and a low-frequency attention matrix. The specific process is as follows:

[0098] ;

[0099] ;

[0100] ;

[0101] ;

[0102] Here, i represents the i-th head in the multi-head attention mechanism, and n represents the total number of heads. The most frequent query value for the i-th head. Let i be the lowest frequency query value of the i-th head. For the key of the i-th head of the r-th sampling rate, Let i be the value of the i-th head of the r-th sampling rate. This is the weight matrix for the most frequent query of the i-th head. Let be the lowest frequency query weight matrix for the i-th head. Let i be the key weight matrix of the i-th head in the r-th sampling rate group. Let be the weight matrix of the i-th head value for the r-th sampling rate group.

[0103] Next, calculate the attention weights:

[0104] ;

[0105] ;

[0106] in, The most frequent attention weight is the i-th one. Let be the lowest frequency attention weight (i-th). Then, the high-frequency cross-attention matrix is ​​calculated using the following formula:

[0107] ;

[0108] The low-frequency cross-attention matrix is ​​calculated using the following formula:

[0109] ;

[0110] in, To project the concatenated result back to the trainable weight matrix of the original dimension, Concat is the concatenation operation. The first superscript 1 and R of Z indicate that the query value corresponds to the first group (highest frequency) and the Rth group (lowest frequency), respectively. The second superscript 1, r, ..., R of Z indicate that the key and value correspond to the first, r, ..., Rth groups, respectively.

[0111] In an optional embodiment, fusing all the bidirectional cross-attention matrices according to the average fusion principle to obtain bidirectional cross-attention features specifically includes:

[0112] The high-frequency attention matrix and low-frequency attention matrix of each scale feature are fused to obtain the fusion matrix of each scale feature;

[0113] Each of the fusion matrices is concatenated, and the concatenated matrices are then dimensionally compressed to obtain the bidirectional cross-attention features.

[0114] Specifically, based on the generated multiple attention matrices, feature fusion is performed according to the average fusion principle to obtain bidirectional cross-attention features. The specific fusion rules are as follows:

[0115] For the two sets of cross-attention matrices obtained, Z 1,1 Z 1,r ,…,Z 1,R and Z R,1 Z R,r ,…,Z R,R The sampling rate group corresponding to the second superscript of Z is fused, assuming it comes from the two attention matrices Z above. 1,r Z R,r ∈R B×1×D The average fusion operator is defined as follows:

[0116] ;

[0117] in, Let r be the fusion matrix of the r-th scale feature. Let be the high-frequency attention matrix for the r-th scale feature. Let be the low-frequency attention matrix for the r-th scale feature.

[0118] In B×1×D, B is the zeroth dimension, representing the batch size; 1 is the first dimension, representing the sequence length after adaptive average pooling; and D is the second dimension, representing the feature dimension, which indicates the length or number of channels of each feature vector.

[0119] According to this fusion rule, the two obtained sets are matched with the sampling rate set by the second superscript, i.e. and r=1,2,…,R corresponds; thus and By fusion And so on, until finally we got , … .

[0120] Next, the obtained , … The concatenation is performed along the feature dimension to obtain the concatenated feature Zconcat. The Zconcat is then input into a linear dimensionality reduction layer to compress the dimensionality and reduce redundant information. This ultimately yields compact cross-attention features. B represents the zeroth dimension, which is the batch size; 1 represents the first dimension, which is the sequence length, and it has been compressed to a single time step; D represents the second dimension, which is the feature dimension, and after dimensionality reduction, it becomes the same size as the feature dimension after the previous convolution.

[0121] Step S104: The historical target value sequence is converted into historical features by the attention module through historical target values, and the bidirectional cross-attention features are dynamically calibrated based on the historical features to obtain a soft measurement model, wherein the historical features have the same dimension as the bidirectional cross-attention features.

[0122] Specifically, the historical target value sequence refers to the true target values ​​of multiple known consecutive time steps in the past of the target value to be predicted at the current time. A historical target value-guided attention module is constructed, and the historical trend information of the target value is used to dynamically calibrate the bidirectional cross-attention features, optimize the cross-attention features, highlight the key features that match the historical trend, and finally construct the soft measurement model.

[0123] In an optional embodiment, before the attention module guides the historical target value sequence to convert it into historical features, the method further includes:

[0124] Based on the samples from the previous preset number of time steps corresponding to the current moment, construct the historical target value sequence for the current moment.

[0125] Specifically, a historical target value sequence Y is constructed based on the data collected in S101. When constructing the historical target value sequence Y, the first preset number of time steps are taken to construct the historical target value sequence Y. Preferably, the preset number of time steps is six time steps, which can avoid the error caused by historical target values ​​that are too old.

[0126] For example, the sample of the historical target value sequence Y constructed in this way at time t is {y} t-6 , y t-5 , ..., y t-1}, denoted as the historical sequence Y t .

[0127] In an optional embodiment, the step of converting the historical target value sequence into historical features through the historical target value-guided attention module, and dynamically calibrating the bidirectional cross-attention features based on the historical features to obtain a soft measurement model, specifically includes:

[0128] The dimension of the bidirectional cross-attention feature is compressed by linear embedding, and the current time-time sequence of historical target values ​​is transformed into historical features with the same dimension as the compressed bidirectional cross-attention feature by linear embedding.

[0129] Calculate the query value corresponding to the historical features based on the historical features;

[0130] Based on the compressed bidirectional cross-attention features, calculate the keys and values ​​of the compressed bidirectional cross-attention features;

[0131] Calculate the attention weight guided by the historical features based on the query value, key, and value.

[0132] Based on the attention weights guided by the historical features, the compressed bidirectional cross-attention features are dynamically calibrated to obtain the feature vector guided by the historical target value sequence, thereby obtaining the soft measurement model.

[0133] Specifically, a linear embedding layer is designed to compress the sequence dimension of A to... Then, the historical target value sequence is transformed into the historical feature with the same dimension as the compressed bidirectional cross-attention feature through linear embedding, ensuring dimension matching during attention computation.

[0134] Next, the historical sequence Y after dimension matching t The value is used as the query value after linear transformation. After linear transformation, they are used as keys and values ​​respectively:

[0135] ;

[0136] Among them, Q H This is a historical sequence query value. It is a historical query weight matrix. yes Dimensions It is the projection dimension of the query or key.

[0137] ;

[0138] ;

[0139] Among them, K H As a key, V H For value, These are the corresponding key weight matrices and value weight matrices. It is the dimension of the feature.

[0140] Calculate the weights using the attention formula:

[0141] ;

[0142] Where, d k W is the feature dimension. H Assign attention weights, and then apply the attention weights:

[0143] ;

[0144] in, This is the feature vector guided by historical target values.

[0145] After obtaining this feature vector guided by historical target values, it is used to predict the output through a fully connected layer, and then the model parameters are optimized end-to-end through backpropagation to finally train this soft measurement model.

[0146] Step S105: Input the data to be predicted in the industrial process into the soft measurement model to obtain the prediction results of the quality variables in the industrial process.

[0147] It should be noted that the data to be predicted in an industrial process refers to the process variable corresponding to the quality variable to be predicted, and the data to be predicted in an industrial process is preferably data from the test set.

[0148] It is worth mentioning that the soft measurement method for industrial multi-rate acquisition provided by this invention collects multi-scale time-series data containing process variables and quality variables. The data is divided into multiple groups according to the sensor sampling frequency and arranged from high to low frequency, laying a data foundation for subsequent multi-scale feature extraction.

[0149] Based on a parallel multi-scale convolutional neural network module, an independent branch-based network architecture is used to extract features from data with different sampling frequencies. Then, by dimensional regularization of the output of each branch, a consistent feature structure is ensured, avoiding the feature distortion problem that may be caused when a single network processes heterogeneous sampled data.

[0150] We designed a bidirectional cross-attention module, using the highest and lowest frequency feature branches as dual query sources, to mine the bidirectional correlation of features across the entire scale, fully covering the dynamic influence and constraints of cross-scale variables.

[0151] By employing the average fusion principle to fuse multiple attention matrices, the association information brought by the most frequent and least frequent queries can be preserved in a balanced manner.

[0152] By guiding the attention module with historical target values, embedding the historical target value sequence into a feature vector of the same dimension, and dynamically calibrating the cross-attention features, historical trend information can be fully utilized, which helps to improve the accuracy of quality prediction in process industries such as oil refining and chemical industry.

[0153] In a specific embodiment, such as Figure 2 As shown, this soft measurement method for industrial multi-rate acquisition can specifically include:

[0154] S1: Collect time-series data from multiple sensors in the process industry system. The data must cover both process variables and quality variables. Based on the sensor sampling frequency, divide the time-series data into multiple groups and arrange them from high to low sampling frequency.

[0155] S2: Construct a parallel multi-scale convolutional neural network module. The module includes multiple independent branches, each branch corresponding to data with different sampling frequencies. Each group of data is used to extract features using a unified structure to obtain multi-scale features of multi-sampling rate data.

[0156] Specifically, a parallel multi-scale convolutional neural network module is constructed. The module includes multiple independent branches, each of which performs independent feature extraction for data with different sampling frequencies, extracts spatiotemporal dynamic features at each time scale, and avoids the problem of time misalignment. Then, the output feature dimensions are kept consistent through dimension mapping, providing a unified feature representation basis for subsequent multi-scale information collaboration.

[0157] S3: Design a bidirectional cross-attention module to achieve bidirectional correlation mining of high-frequency and low-frequency features; then complete feature fusion according to the average fusion principle;

[0158] Specifically, a bidirectional cross-attention module is constructed to achieve bidirectional association mining of high-frequency and low-frequency features. The highest and lowest frequency branch features are used as queries, and all frequency branch features are used as keys and values ​​to calculate the bidirectional cross-attention matrix. Subsequently, based on the generated multiple attention matrices, feature fusion is performed according to the average fusion principle, balancing the retention of association information brought by high-frequency and low-frequency queries.

[0159] S4: Construct a historical target value-guided attention module, introduce historical target value features to dynamically calibrate cross-attention features, and highlight key features that match historical trends;

[0160] Specifically, a historical target value-guided attention module is constructed, which converts the historical target value sequence into a historical feature vector consistent with the dimension of the cross-attention feature through linear embedding; using the historical feature vector as the query and the fused bidirectional cross-attention feature as the key and value, the attention weight guided by the historical target value is calculated and the bidirectional cross-attention feature is dynamically calibrated, thereby highlighting the key features that match the historical trend, thus obtaining a soft measurement model based on multi-scale convolutional bidirectional cross-attention.

[0161] Specifically, such as Figure 3 As shown, Figure 3 The soft measurement model for multi-scale convolutional bidirectional cross-attention includes the following steps: standardization, multi-scale convolutional feature extraction, bidirectional cross-attention and attention feature fusion, historical target value-guided attention, and fully connected layers.

[0162] S5: Input the data from the test set into the prediction model in sequence to obtain the prediction results.

[0163] Specifically, the data in the test set are sequentially input into the soft measurement model based on multi-scale convolutional bidirectional cross-attention to obtain the prediction results.

[0164] In another specific embodiment, the present invention uses a dataset of initial boiling points of heavy naphtha from a hydrocracking process for experimental purposes.

[0165] like Figure 4 As shown, hydrocracking is one of the petroleum refining processes. It is a process in which heavy oil undergoes a cracking reaction under conditions of heating, high hydrogen pressure, and the presence of a catalyst, and is converted into gas, gasoline, jet fuel, diesel, etc.

[0166] In the hydrocracking process, multivariate time-series data within the reactor are collected in real time using industrial sensors. These data cover key indicators in the hydrocracking process, including top bed temperature, bottom bed temperature, reactor differential pressure, tower top pressure, and tower top reflux flow rate, totaling 42 indicators. The mass variable is the initial boiling point of heavy naphtha, as shown in Table 1.

[0167] Table 1: Process Variables and Quality Variables

[0168]

[0169] These indicators comprehensively reflect the operational status of the hydrocracking process, providing rich information for subsequent data analysis and model building. The 42 process variables were divided into three groups based on sampling rate: the first group sampled every 15 minutes (62,400 samples); the second group sampled every 30 minutes (31,200 samples); and the third group sampled every hour (15,600 samples). The mass variables were sampled every 12 hours (1,300 samples). These 1,300 labeled samples were time-aligned with the grouped process variable data according to their sampling timestamps. Subsequently, each aligned mass-labeled sample was integrated with all subsequent process variable observations until the next mass label appeared. Each integrated data unit contains a target label and the context of all process variables within its corresponding time period, collectively forming a complete training sample (i.e., after grouping the 42 process variables, the sampled mass variables were labeled into each group, and then a labeled sample plus the unlabeled process variable samples between the current and next labeled samples and the corresponding target label constituted a training sample; this process was repeated for each group). After the above processing, each group yielded 1300 labeled samples, each containing all unlabeled process variable data for its corresponding time period. These 1300 labeled samples and all unlabeled samples within their respective time periods were then divided into training and test sets according to chronological order at a ratio of 80% and 20%. These process and quality variable data were then used to predict the initial boiling point of heavy naphtha at the current moment. A comparison of the predicted and actual values ​​is shown in the figure below. Figure 5 As shown.

[0170] The soft measurement method for multi-rate data acquisition in industry provided by this invention first collects historical data from the industrial process and divides the data in the training set corresponding to the historical data according to the sampling frequency to obtain multiple sampling rate data. The historical data includes process variables and quality variables. Then, multiple independent branches in the parallel multi-scale convolutional neural network module extract independent features from the sampling rate data at different sampling frequencies, and the outputs after independent feature extraction are dimensionally normalized to obtain scale features for each sampling rate data. Next, using the scale features at the highest and lowest sampling frequencies as query sources, a bidirectional cross-attention module is used to obtain a bidirectional cross-attention matrix for each scale feature based on each scale feature and the query source. All bidirectional cross-attention matrices are then fused according to the average fusion principle to obtain bidirectional cross-attention features. Then, a historical target value-guided attention module converts the historical target value sequence into historical features, and the bidirectional cross-attention features are dynamically calibrated based on the historical features to obtain a soft measurement model. The historical features and the bidirectional cross-attention features have the same dimension. Finally, the data to be predicted in the industrial process is input into the soft measurement model to obtain the prediction results of the quality variables in the industrial process. This invention collects historical data containing process and quality variables, divides the data into multiple sampling rate data based on sampling frequency, laying the data foundation for subsequent multi-scale feature extraction. Through a parallel multi-scale convolutional neural network module, it performs independent feature extraction on data with different sampling frequencies using multiple independent branches. Then, it performs dimensionality regularization on the outputs of each branch to ensure a consistent feature structure, avoiding feature distortion that may occur when a single network processes heterogeneous sampling data. Through a bidirectional cross-attention module, using the highest and lowest frequency feature branches as dual query sources, it mines the bidirectional correlation of full-scale features, obtaining a bidirectional cross-attention matrix for full-scale features, thus ensuring complete... This method covers the dynamic influence and constraints of cross-scale variables and uses the average fusion principle to fuse multiple attention matrices. It can retain the correlation information brought by the most frequent and least frequent queries in a balanced way. By guiding the attention module with historical target values, the historical target value sequence is embedded into a feature vector of the same dimension, and the cross-attention features are dynamically calibrated. This fully utilizes the historical target trend information and ultimately greatly improves the accuracy of industrial system quality prediction. It solves the problem that existing soft measurement methods for industrial systems have insufficient feature extraction, insufficient cross-scale correlation mining, and limited utilization of historical target trend information when processing multi-sampling rate data, resulting in low accuracy of industrial process quality variable prediction.

[0171] Example 2:

[0172] like Figure 6As shown, this embodiment provides a soft measurement device for industrial multi-rate data acquisition, used to execute the above-described soft measurement method for industrial multi-rate data acquisition, including:

[0173] The data acquisition and segmentation module 11 is used to acquire historical data from the industrial process and segment the data in the training set corresponding to the historical data according to the sampling frequency to obtain multiple sampling rate data. The historical data includes process variables and quality variables.

[0174] The extraction and normalization module 12 is connected to the acquisition and division module 11. It is used to extract independent features from the sampling rate data with different sampling frequencies through multiple independent branches in the parallel multi-scale convolutional neural network module, and to normalize the dimensions of the output after independent feature extraction to obtain the scale features of each sampling rate data.

[0175] A fusion module 13 is obtained and connected to the extraction and normalization module 12. It is used to take the scale features with the highest and lowest sampling frequencies as query sources, obtain the bidirectional cross-attention matrix of each scale feature through the bidirectional cross-attention module according to each scale feature and the query source, and fuse all the bidirectional cross-attention matrices according to the average fusion principle to obtain bidirectional cross-attention features.

[0176] The conversion calibration module 14, connected to the fusion module 13, is used to guide the attention module to convert the historical target value sequence into historical features through historical target values, and to dynamically calibrate the bidirectional cross-attention features based on the historical features to obtain a soft measurement model, wherein the historical features have the same dimension as the bidirectional cross-attention features;

[0177] The input module 15, connected to the conversion calibration module 14, is used to input the data to be predicted in the industrial process into the soft measurement model to obtain the prediction results of the quality variables in the industrial process.

[0178] Furthermore, each of the independent branches sequentially comprises: a first 1D convolutional layer, a first batch normalization layer, a first ReLU activation function, a max pooling layer, a second 1D convolutional layer, a second batch normalization layer, a second ReLU activation function, and an adaptive average pooling layer.

[0179] Furthermore, the fusion module 13 specifically includes:

[0180] A sorting unit is used to sort the scale features of all the sampling rate data from high to low according to the sampling frequency;

[0181] The first calculation unit is used to calculate the highest frequency query value based on the scale feature of the highest sampling frequency;

[0182] The second calculation unit is used to calculate the lowest frequency query value based on the scale feature of the lowest sampling frequency;

[0183] The third calculation unit is used to calculate the key and value of each scale feature based on each scale feature;

[0184] The fourth calculation unit is used to calculate the bidirectional cross-attention matrix of each scale feature using a multi-head attention mechanism, based on the highest frequency query value, the lowest frequency query value, and the key and value of each scale feature.

[0185] Furthermore, the bidirectional cross-attention matrix includes a high-frequency attention matrix and a low-frequency attention matrix, and the fusion module 13 further includes:

[0186] The fusion unit is used to fuse the high-frequency attention matrix and the low-frequency attention matrix of each scale feature to obtain the fusion matrix of each scale feature.

[0187] The splicing and compression unit is used to splice each of the fusion matrices and compress the dimensions of the spliced ​​matrices to obtain the bidirectional cross-attention features.

[0188] Furthermore, the fusion-derived unit specifically includes:

[0189] The high-frequency attention matrix and low-frequency attention matrix of each scale feature are fused using the following formula:

[0190] ;

[0191] in, Let r be the fusion matrix of the r-th scale feature. Let be the high-frequency attention matrix for the r-th scale feature. Let be the low-frequency attention matrix for the r-th scale feature.

[0192] Furthermore, the device also includes:

[0193] The construction unit is used to construct the historical target value sequence for the current moment based on the samples of the previous preset number of time steps corresponding to the current moment.

[0194] Furthermore, the conversion calibration module 14 specifically includes:

[0195] The compression and transformation unit is used to compress the dimension of the bidirectional cross-attention feature through linear embedding, and to transform the current historical target value sequence into the historical feature with the same dimension as the compressed bidirectional cross-attention feature through linear embedding.

[0196] The fifth calculation unit is used to calculate the query value corresponding to the historical features based on the historical features;

[0197] The sixth calculation unit is used to calculate the keys and values ​​of the compressed bidirectional cross-attention features based on the compressed bidirectional cross-attention features;

[0198] The seventh calculation unit is used to calculate the attention weight guided by the historical features based on the query value, key, and value.

[0199] The calibration unit is used to dynamically calibrate the compressed bidirectional cross-attention features based on the attention weights guided by the historical features, to obtain the feature vector guided by the historical target value sequence, so as to obtain the soft measurement model.

[0200] Example 3:

[0201] refer to Figure 7 This embodiment provides a soft measurement device for industrial multi-rate acquisition, including a memory 21 and a processor 22. The memory 21 stores a computer program, and the processor 22 is configured to run the computer program to execute the soft measurement method for industrial multi-rate acquisition in Embodiment 1.

[0202] The memory 21 is connected to the processor 22. The memory 21 can be a flash memory, a read-only memory or other memory, and the processor 22 can be a central processing unit or a microcontroller.

[0203] Example 4:

[0204] This embodiment provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the soft measurement method for industrial multi-rate acquisition described in Embodiment 1 above.

[0205] The computer-readable storage medium includes volatile or non-volatile, removable or non-removable media implemented in any method or technology for storing information, such as computer-readable instructions, data structures, computer program modules or other data. Computer-readable storage media include, but are not limited to, RAM (Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other memory technologies, CD-ROM (Compact Disc Read-Only Memory), DVD or other optical disc storage, cartridges, magnetic tapes, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer.

[0206] In summary, the soft measurement method, device, and medium for industrial multi-rate data acquisition provided in this embodiment of the invention first collects historical data from industrial processes and divides the data in the training set corresponding to the historical data according to the sampling frequency to obtain multiple sampling rate data. The historical data includes process variables and quality variables. Then, multiple independent branches in a parallel multi-scale convolutional neural network module extract independent features from the sampling rate data at different sampling frequencies, and the outputs after independent feature extraction are dimensionality-normalized to obtain the scale features of each sampling rate data. Finally, using the scale features of the highest and lowest sampling frequencies as the query source, according to... For each scale feature and the query source, a bidirectional cross-attention matrix for each scale feature is obtained through a bidirectional cross-attention module. All bidirectional cross-attention matrices are then fused according to an average fusion principle to obtain a bidirectional cross-attention feature. Next, a historical target value-guided attention module converts the historical target value sequence into historical features, and the bidirectional cross-attention feature is dynamically calibrated based on these historical features to obtain a soft measurement model. The historical features and the bidirectional cross-attention feature have the same dimension. Finally, the data to be predicted in the industrial process is input into the soft measurement model to obtain the prediction results for the quality variables in the industrial process. This invention collects historical data containing process and quality variables, divides the data into multiple sampling rate data based on sampling frequency, laying the data foundation for subsequent multi-scale feature extraction. Through a parallel multi-scale convolutional neural network module, it performs independent feature extraction on data with different sampling frequencies using multiple independent branches. Then, it performs dimensionality regularization on the outputs of each branch to ensure a consistent feature structure, avoiding feature distortion that may occur when a single network processes heterogeneous sampling data. Through a bidirectional cross-attention module, using the highest and lowest frequency feature branches as dual query sources, it mines the bidirectional correlation of full-scale features, obtaining a bidirectional cross-attention matrix for full-scale features, thus ensuring complete... This method covers the dynamic influence and constraints of cross-scale variables and uses the average fusion principle to fuse multiple attention matrices. It can retain the correlation information brought by the most frequent and least frequent queries in a balanced way. By guiding the attention module with historical target values, the historical target value sequence is embedded into a feature vector of the same dimension, and the cross-attention features are dynamically calibrated. This fully utilizes the historical target trend information and ultimately greatly improves the accuracy of industrial system quality prediction. It solves the problem that existing soft measurement methods for industrial systems have insufficient feature extraction, insufficient cross-scale correlation mining, and limited utilization of historical target trend information when processing multi-sampling rate data, resulting in low accuracy of industrial process quality variable prediction.

[0207] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A soft measurement method for industrial multi-rate data acquisition, characterized in that, The method includes: Historical data from industrial processes are collected, and the data in the training set corresponding to the historical data are divided according to the sampling frequency to obtain multiple sampling rate data. The historical data includes process variables and quality variables. The sampling rate data at different sampling frequencies are extracted independently by multiple independent branches in the parallel multi-scale convolutional neural network module, and the output after independent feature extraction is dimensionally normalized to obtain the scale features of each sampling rate data. Using the scale features with the highest and lowest sampling frequencies as query sources, and based on each scale feature and the query source, a bidirectional cross-attention matrix for each scale feature is obtained through a bidirectional cross-attention module. All bidirectional cross-attention matrices are then fused according to the average fusion principle to obtain bidirectional cross-attention features. The historical target value sequence is converted into historical features by guiding the attention module through historical target values, and the bidirectional cross-attention features are dynamically calibrated based on the historical features to obtain a soft measurement model, wherein the historical features have the same dimension as the bidirectional cross-attention features; The data to be predicted in the industrial process is input into the soft measurement model to obtain the prediction results of the quality variables in the industrial process.

2. The method according to claim 1, characterized in that, Each of the independent branches sequentially comprises: a first 1D convolutional layer, a first batch normalization layer, a first ReLU activation function, a max pooling layer, a second 1D convolutional layer, a second batch normalization layer, a second ReLU activation function, and an adaptive average pooling layer.

3. The method according to claim 1, characterized in that, The process involves using the scale features with the highest and lowest sampling frequencies as query sources, and obtaining a bidirectional cross-attention matrix for each scale feature through a bidirectional cross-attention module based on each scale feature and the query source. Specifically, this includes: The scale features of all the sampling rate data are sorted from high to low according to the sampling frequency; Calculate the highest frequency query value based on the scale feature of the highest sampling frequency; Calculate the lowest frequency query value based on the scale characteristics of the lowest sampling frequency; Calculate the key and value for each scale feature based on each scale feature; Based on the highest and lowest frequency query values, as well as the key and value of each scale feature, a multi-head attention mechanism is used to calculate the bidirectional cross-attention matrix for each scale feature.

4. The method according to claim 3, characterized in that, The bidirectional cross-attention matrix includes a high-frequency attention matrix and a low-frequency attention matrix. The process of fusing all the bidirectional cross-attention matrices according to an average fusion principle to obtain bidirectional cross-attention features specifically includes: The high-frequency attention matrix and low-frequency attention matrix of each scale feature are fused to obtain the fusion matrix of each scale feature; Each of the fusion matrices is concatenated, and the concatenated matrices are then dimensionally compressed to obtain the bidirectional cross-attention features.

5. The method according to claim 4, characterized in that, The process of fusing the high-frequency attention matrix and the low-frequency attention matrix for each scale feature to obtain a fusion matrix for each scale feature specifically includes: The high-frequency attention matrix and low-frequency attention matrix of each scale feature are fused using the following formula: ； in, Let r be the fusion matrix of the r-th scale feature. Let be the high-frequency attention matrix for the r-th scale feature. Let be the low-frequency attention matrix for the r-th scale feature.

6. The method according to claim 1, characterized in that, Before the attention module, guided by historical target values, converts the sequence of historical target values ​​into historical features, the method further includes: Based on the samples from the previous preset number of time steps corresponding to the current moment, construct the historical target value sequence for the current moment.

7. The method according to claim 6, characterized in that, The step of converting the historical target value sequence into historical features through the historical target value-guided attention module, and dynamically calibrating the bidirectional cross-attention features based on the historical features to obtain a soft measurement model, specifically includes: The dimension of the bidirectional cross-attention feature is compressed by linear embedding, and the current time-time sequence of historical target values ​​is transformed into historical features with the same dimension as the compressed bidirectional cross-attention feature by linear embedding. Calculate the query value corresponding to the historical features based on the historical features; Based on the compressed bidirectional cross-attention features, calculate the keys and values ​​of the compressed bidirectional cross-attention features; Calculate the attention weight guided by the historical features based on the query value, key, and value. Based on the attention weights guided by the historical features, the compressed bidirectional cross-attention features are dynamically calibrated to obtain the feature vector guided by the historical target value sequence, thereby obtaining the soft measurement model.

8. A soft measurement device for industrial multi-rate data acquisition, characterized in that, include: The data acquisition and segmentation module is used to collect historical data from industrial processes and segment the data in the training set corresponding to the historical data according to the sampling frequency to obtain multiple sampling rate data. The historical data includes process variables and quality variables. An extraction and normalization module, connected to the acquisition and partitioning module, is used to extract independent features from the sampling rate data at different sampling frequencies through multiple independent branches in the parallel multi-scale convolutional neural network module, and to normalize the dimensions of the output after independent feature extraction to obtain the scale features of each sampling rate data. A fusion module is obtained and connected to the extraction and normalization module. It is used to take the scale features with the highest and lowest sampling frequencies as query sources, and obtain the bidirectional cross-attention matrix of each scale feature through the bidirectional cross-attention module according to each scale feature and the query source. Then, it fuses all the bidirectional cross-attention matrices according to the average fusion principle to obtain the bidirectional cross-attention feature. A conversion calibration module, connected to the fusion module, is used to guide the attention module to convert the historical target value sequence into historical features through historical target values, and to dynamically calibrate the bidirectional cross-attention features based on the historical features to obtain a soft measurement model, wherein the historical features have the same dimension as the bidirectional cross-attention features; The input module, connected to the conversion calibration module, is used to input the data to be predicted in the industrial process into the soft measurement model to obtain the prediction results of the quality variables in the industrial process.

9. A soft measurement device for industrial multi-rate data acquisition, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to implement the soft measurement method for industrial multi-rate acquisition as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the soft measurement method for industrial multi-rate acquisition as described in any one of claims 1-7.

Images